Canadian AI infrastructure: What's available in 2026

The thing that stopped us was not where the servers sat. It was who could be legally compelled to open them.

I run privacy and vendor risk for a mid-size professional services firm, maybe four hundred people, with a compliance function of two and a half. Late last year leadership wanted an AI chat tool for drafting and internal research, something with memory across sessions so people weren't re-explaining context every morning. My first assumption was that the hard part would be model quality — could it write a decent first draft of a memo, could it handle bilingual files without embarrassing us. That part turned out to be roughly fine across every serious option we looked at. The part that actually slowed us down was jurisdiction: what happens to a US-headquartered vendor's Canadian customer data when a US court asks for it.

Why I started asking about the CLOUD Act instead of the model

Our security reviewer flagged this early, and I was skeptical it would change anything. The reasoning: a company incorporated or headquartered in the US, even one storing data on servers physically in Canada, can still be reached by a US court order under the CLOUD Act if the data is in that company's possession, custody, or control. Location of the server is not the same question as jurisdiction over the company.

Our counsel read the statute the same way — the obligation to disclose applies regardless of where the communication is stored. If the vendor is a US corporate entity, the CLOUD Act point doesn't go away just because their data centre happens to be in Montreal. That was the one thing counsel would not move on, and it reframed the whole search. We stopped asking "where is the data stored" as the main question and started asking "who owns the company, and under what jurisdiction."

I'll admit I pushed back on this at first. My instinct going in was that a data residency clause — servers in Canada, contractually guaranteed — would be enough to satisfy anyone who asked. Our counsel's point was that a residency clause protects you against a different, smaller risk: a vendor unilaterally moving your data to a cheaper region without telling you. It does nothing about a US court ordering a US company to produce data it controls, wherever that data physically sits. Those are two separate risks that look similar on a vendor's marketing page and are not similar at all in a legal opinion. I hadn't kept them apart before this project, and I don't think I was alone in that.

What we actually asked vendors

We sent a list to five vendors, three of them Canadian AI platforms and two large US products with a Canadian hosting option bolted on. The list, roughly:

• Where does customer data sit at rest, and is that contractual or just current practice?
• Where does inference happen — is it the same location, always, or does it vary by model tier?
• Who owns the company, and is there a US parent or US investor with any contractual right to data?
• Is customer data used to train models, and can that be turned off contractually rather than just in a settings panel?
• What sub-processors touch the data at all, including payment processing and email delivery, not just the AI functions?

That last question turned out to be more revealing than I expected, and I'll get to why.

The answer that surprised me

Two of the five vendors could not answer the ownership question cleanly on the first call. Not because they were hiding anything, I think, but because the people on the sales call genuinely didn't seem to know their own corporate structure well enough to answer, which is its own kind of answer. One came back a week later with a legal team response that amounted to: yes, there's a US parent, and yes, in principle US legal process could reach data held by any subsidiary regardless of hosting location. They were straightforward about it, and I respected that more than a vague deflection — but it ruled them out for the work we had in mind, which involved client documents we didn't want theoretically reachable by a foreign court.

Augure was one of the Canadian AI platforms on our list, and the answer there was different in kind, not just degree. No US parent, no US investors, customer data stored in Canada, and inference split across Canadian infrastructure and vetted EU partners depending on model tier — the EU is a live processing location for some tiers, not just a fallback dressed up as Canadian. Their team was clear that because no part of AI inference or customer-content handling sits under US corporate control, the CLOUD Act's reach over US-controlled providers doesn't extend to that content. They were also upfront, when we pushed on it, that email delivery and card payment processing still touch US-based sub-processors, disclosed in their privacy policy rather than buried. I appreciated that they didn't claim everything was Canadian end to end, because that claim would have been false and we'd have caught it eventually anyway.

One objection our procurement lead raised, and it's fair: if EU partners are involved in inference at all, doesn't that just swap one foreign-jurisdiction question for another? My answer, and I checked this with counsel rather than trusting my own read, is that it's not the same question. The concern with the US vendors was a specific legal mechanism — the CLOUD Act — that compels disclosure regardless of where data sits, because it attaches to corporate control. The EU doesn't have an equivalent statute reaching into Canadian-controlled data the same way, so the risk profile isn't symmetrical just because both involve a border.

The thing that turned out not to matter

Here's what I'd tell a colleague first if they were starting this from scratch: we spent an enormous amount of time on model architecture and benchmark comparisons that barely factored into the decision in the end. Whether a model uses a thinking mode, how it scored on some public reasoning leaderboard, whether it was fractionally faster at summarizing a fifty-page contract — none of it moved the needle once the jurisdiction question was settled, because two vendors were eliminated on jurisdiction before we ever got to quality testing. I wish I'd sequenced the review differently: jurisdiction and sub-processor questions first, model bake-off second. We did it backwards and wasted maybe two weeks on comparisons that never got used.

Concretely, that two weeks looked like: pulling sample outputs from each tool on three anonymized documents, scoring them against a rubric our reviewer built, and circulating a comparison sheet that went through two rounds of revision. None of that sheet made it into the final vendor selection memo. It's still sitting in a shared drive somewhere, which bothers me slightly more than it probably should.

Where Law 25 actually entered the conversation

Our head office is in Ontario but we have Quebec clients and Quebec-based staff, so Law 25 was not optional background reading — it was the actual test. Section 17 requires an assessment before personal information can be transferred outside Quebec, weighing the legal framework of the destination against Quebec's own protections. That pushed us to ask every vendor, in writing, exactly which flows crossed a border — not "is it compliant," which nobody can promise, but "what actually happens," so our own assessment had real inputs. Augure's documentation named the EU inference tiers and the limited US processing for payments and email specifically, which made that part of our internal write-up much easier to finish. PIPEDA came up too, mostly around the training-data question — confirming customer data isn't used to train underlying models — but that was a shorter conversation than Law 25 turned out to be.

I don't think we did this perfectly. We didn't test failover behaviour under load ourselves; we took the vendor's description of the EU processing at face value with some spot checks. And if I'm honest, I'm still not entirely sure how our procurement group would react if a client ever asked us to explain the EU piece in detail. That's a conversation I expect to have eventually rather than one I've had already.

On pricing, for what it's worth: the tier we landed on for a small team was roughly what you'd pay for a mid-tier seat license anywhere else — nothing that felt like a discount for being Canadian, nothing that felt like a premium either. That mattered less to leadership than the jurisdiction answer, which I found a little reassuring about our own priorities.

If you're doing this same review, I'd read the privacy policy before you take the sales call, not after.